Medical image registration method and device, electronic device, and storage medium

Abstract

The embodiments of the present application provide a medical image registration method and device, electronic device, and storage medium, which relate to the field of artificial intelligence technology. The method includes: obtaining a target floating medical image and a target reference medical image; inputting the target floating medical image and the target reference medical image into a medical image registration model to obtain a target deformation field, thereby obtaining a target registered medical image; wherein the medical image registration model includes an encoding module, a third-order transformation convolution submodule, a convolution attention submodule, a fusion convolution submodule, and a decoding module, and is trained based on a first loss function and a second loss function, the first loss function is constructed based on the distance between the sample reference medical image and the sample registered medical image, and the second loss function is constructed based on the distance between the sample deformation field and a preset identical deformation field. The present application improves the accuracy of medical image registration by simultaneously optimizing the model structure and the loss function.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a medical image registration method and a medical image registration device, an electronic device, and a storage medium. Background Art

[0002] Artificial intelligence technology is rapidly developing and is widely used in daily production and life. In particular, in the medical field, numerous computer-assisted diagnosis methods and approaches have emerged. Medical image registration is a key technology in medical image processing. It involves spatially aligning one or more medical images with another to achieve anatomical consistency.

[0003] There is still much room for improvement in the accuracy of existing algorithms. Summary of the Invention

[0004] The main purpose of the embodiments of the present application is to propose a medical image registration method and device, electronic equipment, and storage medium, thereby improving the accuracy of medical image registration.

[0005] To achieve the above objectives, a first aspect of an embodiment of the present application provides a medical image registration method, the method comprising:

[0006] acquiring a target floating medical image and a target reference medical image;

[0007] Acquire a medical image registration model, wherein the medical image registration model includes an encoding module, a third-order transform convolution submodule, a convolution attention submodule, a fusion convolution submodule, and a decoding module;

[0008] Inputting the target floating medical image and the target reference medical image into the encoding module for feature encoding to obtain target encoding features;

[0009] Inputting the target coding feature into the third-order transform convolution submodule for feature transformation to obtain the target transformed feature;

[0010] Inputting the target transformation feature into the convolutional attention submodule for feature extraction to obtain the target attention feature;

[0011] Inputting the target attention feature into the fusion convolution submodule for feature convolution to obtain the target fusion feature;

[0012] Inputting the target coding feature and the target fusion feature into the decoding module for feature decoding to obtain a target deformation field, and applying the target deformation field to the target floating medical image to obtain a target registered medical image;

[0013] In which, the medical image registration model is trained based on the first loss function and the second loss function of each sample image pair in the sample image set, the sample image pair includes a sample floating medical image and a sample reference medical image, the first loss function is constructed based on the distance between the sample reference medical image and the sample registered medical image, and the second loss function is constructed based on the distance between the sample deformation field and the preset identical deformation field. The sample deformation field is obtained by inputting the sample floating medical image and the sample reference medical image into the medical image registration model, and the sample registered medical image is obtained by applying the sample deformation field to the sample floating medical image, and each element of the identical deformation field is zero.

[0014] Optionally, the third-order transform convolution submodule includes a third-order transform operator layer, a first convolution layer, a normalization layer, and an activation function layer; the step of inputting the target coding feature into the third-order transform convolution submodule for feature transformation to obtain the target transformed feature includes:

[0015] Using the third-order transformation operator layer, performing a third-order feature transformation on the target coding feature to obtain a first transformation feature;

[0016] Convolving the first transformed features using the first convolutional layer to obtain second transformed features;

[0017] Normalizing the second transformed features using the normalization layer to obtain third transformed features;

[0018] The activation function layer is used to perform nonlinear processing on the third transformation feature to obtain the target transformation feature.

[0019] Optionally, the convolutional attention submodule includes a second convolutional layer, an attention layer, an upsampling layer, and a residual layer; and inputting the target transformation feature into the convolutional attention submodule for feature extraction to obtain the target attention feature includes:

[0020] Using the second convolutional layer, convolving the target transformation feature to obtain a first extracted feature;

[0021] Using the attention layer, performing attention processing on the first extracted features to obtain second extracted features;

[0022] Upsampling the second extracted features using the upsampling layer to obtain third extracted features;

[0023] The target transformation feature and the third extracted feature are concatenated using the residual layer to obtain the target attention feature.

[0024] Optionally, the fused convolution submodule includes a plurality of cascaded dilated convolution layers; and inputting the target attention feature into the fused convolution submodule to obtain the target fused feature includes:

[0025] The target attention feature is used as the convolution input feature of multiple cascaded dilated convolutional layers, so that each dilated convolutional layer after the first dilated convolutional layer uses the convolution output feature of the previous dilated convolutional layer as the convolution input feature, and each dilated convolutional layer performs dilated convolution on the convolution input feature to obtain the convolution output feature;

[0026] The convolution output feature output by the last spatiotemporal convolution layer is used as the target fusion feature.

[0027] Optionally, the encoding module is a cascade of multiple encoding modules, and the target encoding feature includes the encoding output feature of each encoding module; inputting the target floating medical image and the target reference medical image into the encoding module for feature encoding to obtain the target encoding feature includes:

[0028] performing image stitching on the target floating medical image and the target reference medical image to obtain a target stitched medical image;

[0029] The encoding input features of the first encoding module in the cascaded multiple encoding modules are determined according to the target spliced ​​medical image, so that each encoding module after the first encoding module uses the encoding output features of the previous encoding module after maximum pooling processing as the encoding input features, and each encoding module performs feature encoding according to the encoding input features to obtain the encoding output features.

[0030] Optionally, the decoding module is a cascade of multiple decoding modules, one decoding module corresponds to one encoding module; the target encoding feature and the target attention feature are input into the decoding module for feature decoding to obtain a target deformation field, including:

[0031] Perform feature splicing on the output feature of the last encoding module and the target attention feature to obtain a target splicing feature;

[0032] Using the target splicing features as decoding input features of a plurality of cascaded decoding modules, so that each decoding module after the first decoding module uses the decoding output features of the previous decoding module after upsampling processing and the encoding output features of the encoding module corresponding to the decoding module as decoding input features, and each decoding module performs feature decoding according to the decoding input features to obtain the decoding output features;

[0033] The target deformation field is determined according to the decoding output feature of the last decoding module.

[0034] Optionally, the medical image registration model is trained based on a first loss function and a second loss function for each sample image pair in the sample image set, including:

[0035] Calculating a mean square error between the sample reference medical image and the sample registered medical image to obtain a first loss function;

[0036] Calculate the square of the difference between each element in the sample deformation field and the corresponding element in the identity deformation field to obtain a second sub-loss function, and sum the second sub-loss functions of all elements to obtain a second loss function;

[0037] Obtain a target loss function by performing a weighted summation on the first loss function and the second loss function;

[0038] Parameters of the medical image registration model are adjusted according to the target loss function.

[0039] To achieve the above-mentioned purpose, a second aspect of an embodiment of the present application provides a medical image registration device, comprising:

[0040] an image acquisition unit, configured to acquire a target floating medical image and a target reference medical image;

[0041] A model acquisition unit, configured to acquire a medical image registration model, wherein the medical image registration model comprises an encoding module, a third-order transform convolution submodule, a convolution attention submodule, a fusion convolution submodule, and a decoding module;

[0042] A feature encoding unit, configured to input the target floating medical image and the target reference medical image into the encoding module for feature encoding to obtain target encoding features;

[0043] A feature conversion unit, configured to input the target coding feature into the third-order transform convolution submodule for feature conversion to obtain a target transformed feature;

[0044] A feature extraction unit, configured to input the target transformation feature into the convolutional attention submodule for feature extraction to obtain a target attention feature;

[0045] A feature convolution unit, configured to input the target attention feature into the fusion convolution submodule for feature convolution to obtain a target fusion feature;

[0046] a feature decoding unit, configured to input the target coding feature and the target fusion feature into the decoding module for feature decoding to obtain a target deformation field, and apply the target deformation field to the target floating medical image to obtain a target registered medical image;

[0047] In which, the medical image registration model is trained based on the first loss function and the second loss function of each sample image pair in the sample image set, the sample image pair includes a sample floating medical image and a sample reference medical image, the first loss function is constructed based on the distance between the sample reference medical image and the sample registered medical image, and the second loss function is constructed based on the distance between the sample deformation field and the preset identical deformation field. The sample deformation field is obtained by inputting the sample floating medical image and the sample reference medical image into the medical image registration model, and the sample registered medical image is obtained by applying the sample deformation field to the sample floating medical image, and each element of the identical deformation field is zero.

[0048] To achieve the above-mentioned purpose, the third aspect of an embodiment of the present application proposes an electronic device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the method described in the first aspect when executing the computer program.

[0049] To achieve the above-mentioned purpose, the fourth aspect of an embodiment of the present application proposes a storage medium, which is a computer-readable storage medium and stores a computer program. When the computer program is executed by a processor, the method described in the first aspect is implemented.

[0050] The present application proposes a medical image registration method and device, electronic device, and storage medium. The method sets a cascaded third-order transformation convolution module, a convolution attention module, and a fusion convolution module between the encoding module and the decoding module in the medical image registration model, so that the target fusion feature obtained has relatively rich feature information. Then, the decoding and encoding performs feature decoding based on the target fusion feature and the target encoding feature output by the encoding module, and the target deformation field obtained has a high accuracy, so that the target registration medical image obtained based on the target deformation field has a higher accuracy. In addition, in addition to setting the model structure of the medical registration model, the present application also uses the first loss function and the second loss function to train the medical image registration model, thereby further improving the accuracy of the medical registration model. It can be seen that the present application improves the model structure and the model loss function at the same time, so that the medical image registration model has a powerful feature extraction capability, thereby improving the accuracy of the medical image registration. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1This is an optional flowchart of the medical image registration method provided in the embodiment of the present application;

[0052] Figure 2 is a schematic structural diagram of the medical image registration model provided in an embodiment of the present application;

[0053] Figure 3 Schematic diagram of the structure of the third-order transform convolution submodule provided in an embodiment of the present application;

[0054] Figure 4 Schematic diagram of the structure of the convolutional attention submodule provided in an embodiment of the present application;

[0055] Figure 5 Schematic diagram of the structure of the fused convolution submodule provided in the embodiment of the present application;

[0056] Figure 6 is an optional flow chart of a medical image registration method provided by another embodiment of the present application;

[0057] Figure 7 This is a visualization example diagram provided by the embodiment of the present application;

[0058] Figure 8 is a structural diagram of a medical image registration model provided by another embodiment of the present application;

[0059] Figure 9 is a structural diagram of a medical image registration model provided by another embodiment of the present application;

[0060] Figure 10 A schematic diagram of the structure of a medical image registration device provided in an embodiment of the present application;

[0061] Figure 11 This is a schematic diagram of the hardware structure of the electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0062] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0063] It should be noted that although the device schematics illustrate functional module divisions and the flowcharts illustrate logical sequences, in certain circumstances, the steps shown or described may be performed in a sequence that differs from the module divisions in the device or the sequence in the flowcharts. The terms "first," "second," and so on, in the specification, claims, and drawings, are used to distinguish similar items and are not necessarily used to describe a specific sequence or precedence.

[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0065] First, let’s analyze some of the terms used in this application:

[0066] Artificial intelligence (AI) is a new technical discipline that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence. A branch of computer science, AI seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thinking. It also encompasses the theories, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results.

[0067] Back propagation: The general principle of back propagation is: input the training set data into the input layer of the neural network, pass through the hidden layer of the neural network, and finally reach the output layer of the neural network and output the result; since there is an error between the output result of the neural network and the actual result, the error between the estimated value and the actual value is calculated, and the error is backpropagated from the output layer to the hidden layer until it propagates to the input layer; in the process of back propagation, the values ​​of various parameters are adjusted according to the error; the above process is continuously iterated until convergence.

[0068] Image registration: Given a floating image and a reference image, predict a deformation field, i.e., a mapping from the floating image to the reference image, so that the registered floating image and the reference image are as similar as possible.

[0069] Artificial intelligence technology is rapidly developing and is widely used in daily life and production. This is particularly true in the medical field, where numerous computer-assisted diagnosis methods and approaches have emerged. Medical image registration is a key technology in medical image processing. It involves spatially aligning one or more medical images with another to achieve anatomical consistency. This process is crucial for a variety of medical applications. In clinical diagnosis and treatment planning, doctors may need to combine different types of medical images, such as CT, MRI, and PET, to obtain more comprehensive patient information. Image registration aligns these images, generated from different imaging techniques, at different time points, or from different perspectives, to better understand the patient's pathological condition. In tumor diagnosis and testing, registration technology allows for comparison of images of the same tissue from the same patient at different time points to monitor tumor growth, shrinkage, or metastasis, thereby assessing treatment efficacy. During surgery, image registration aligns preoperative medical images with real-time intraoperative images, helping surgeons more precisely locate the surgical area and improving surgical success and safety. However, existing algorithms still have significant room for improvement in registration accuracy and timeliness, and also struggle to achieve satisfactory results with complex images.

[0070] In view of the above problems, the embodiment of the present application aims to provide a medical image registration method, which can improve the accuracy of medical image registration.

[0071] The medical image registration method of the embodiment of the present application can be executed by the server alone, or by the terminal alone, or by the terminal and the server together. In addition, the medical image registration method provided by the embodiment of the present application can also be software running on the server. The server can be configured as an independent physical server, or as a server cluster or distributed system composed of multiple physical servers, or as a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application that implements the medical image registration method, etc., but is not limited to the above forms.

[0072] The embodiments of the present application provide a medical image registration method, a medical image registration device, an electronic device, and a computer-readable storage medium, which are specifically illustrated by the following embodiments. First, the medical image registration method in the embodiments of the present application is described.

[0073] See Figure 1 According to one embodiment of the present application, a medical image registration method includes:

[0074] Step 101, acquiring a target floating medical image and a target reference medical image;

[0075] Step 102: Acquire a medical image registration model, the medical image registration model including an encoding module, a third-order transform convolution submodule, a convolution attention submodule, a fusion convolution submodule, and a decoding module;

[0076] Step 103: Input the target floating medical image and the target reference medical image into a coding module for feature coding to obtain target coding features;

[0077] Step 104: input the target coding feature into the third-order transform convolution submodule for feature transformation to obtain the target transformed feature;

[0078] Step 105: Input the target transformation feature into the convolutional attention submodule for feature extraction to obtain the target attention feature;

[0079] Step 106: Input the target attention feature into the fusion convolution submodule for feature convolution to obtain the target fusion feature;

[0080] In step 107 , the target coding features and the target fusion features are input into a decoding module for feature decoding to obtain a target deformation field, and the target deformation field is applied to the target floating medical image to obtain a target registered medical image.

[0081] Through steps 101 to 107, by redesigning the model backbone network and adopting a convolutional attention submodule that combines convolution and attention mechanisms, the model can better learn the relevant features of the images before and after registration. The use of a third-order transform convolution submodule gives the model powerful feature extraction capabilities and can accurately register tissue regions in medical images. The registration method provided in this application is suitable for assisting clinical disease diagnosis and treatment, improving doctors' diagnostic efficiency.

[0082] In step 101 of some embodiments, the target floating medical image and the target reference medical image are two frames of images of the same target object, where the target object can be a part of the human body, such as the chest. For example, the target floating medical image and the target reference medical image are two frames of images of the same target object at the same time using different imaging modalities or different imaging angles. Of the two frames, one is the target floating medical image, i.e., the image to be transformed, and the other is the target reference medical image, i.e., the image that provides a reference for the transformation of the target floating medical image. For example, the target floating medical image and the target reference medical image can be a pair of magnetic resonance imaging images.

[0083] There are various ways to obtain the target floating medical image and the target reference medical image. For example, various image acquisition devices (such as CT) can be used to acquire images of the target object (e.g., the ventricle) to obtain the target floating medical image and the target reference medical image. Alternatively, the target floating medical image and the target reference medical image can be obtained from a local or external database, or they can be searched online.

[0084] In step 102 of some embodiments, the medical image registration model refers to a model used for image registration. Figure 2 The medical image registration model includes an encoding module, a third-order transformation convolution submodule, a convolution attention submodule, a fusion convolution submodule, and a decoding module, which are connected in sequence, and the decoding module is connected to the encoding module.

[0085] The encoding module refers to a module that performs feature encoding on the target floating medical image and the target reference medical image. The third-order transform convolution submodule refers to a module that performs third-order transform and convolution operations on image features. The convolution attention submodule refers to a module that performs convolution operations and attention transforms on image features. The fusion convolution submodule refers to a module that performs multiple convolution operations on image features. The decoding module refers to a module that performs feature decoding on image features. The present application introduces a third-order transform convolution submodule, a convolution attention submodule, and a fusion convolution submodule between the encoding module and the decoding module, which greatly improves the model's ability to extract features and improves the accuracy of medical image registration.

[0086] In one embodiment, referring to Figure 3 The third-order transform convolution submodule includes a third-order transform operator layer, a first convolution layer, a normalization layer, and an activation function layer. The third-order transform operator layer is used to perform third-order transforms. The first convolution layer is used for convolution. The normalization layer is used for normalization. The activation function layer is used for nonlinear activation. It should be noted that the third-order transform specifically refers to transforming each pixel in the image through a cubic polynomial (also known as a third-order transform operator). The activation function layer can use activation functions such as SiLU and ReLU.

[0087] In one embodiment, referring to Figure 4 The convolutional attention submodule consists of a second convolutional layer, an attention layer, an upsampling layer, and a residual layer. The second convolutional layer performs convolution. The attention layer uses the attention mechanism to perform feature fusion. The upsampling layer performs upsampling. The residual layer concatenates the features input to the second convolutional layer with the features output by the upsampling layer.

[0088] In one embodiment, referring to Figure 5 , the fused convolution submodule consists of multiple cascaded dilated convolution layers. Figure 5Four dilated convolutional layers are shown, but the number of dilated convolutional layers can be adjusted as needed. Dilated convolutional layers are used to perform dilated convolutions. It should be noted that, given the variable shapes of objects to be predicted in medical images, multiple dilated convolutions are used to perform the convolution operation, resulting in an isotropic circular effective receptive field for the fused convolution module.

[0089] In one embodiment, a medical image registration model is trained based on a first loss function and a second loss function for each sample image pair in a sample image set, where the sample image pair includes a sample floating medical image and a sample reference medical image. The first loss function is constructed based on the distance between the sample reference medical image and the sample registered medical image, and the second loss function is constructed based on the distance between the sample deformation field and the identity deformation field. The sample deformation field is obtained by inputting the sample floating medical image and the sample reference medical image into the medical image registration model, and the sample registered medical image is obtained by applying the sample deformation field to the sample floating medical image, and each element of the identity deformation field is zero.

[0090] The sample image set is a data set to be registered, including multiple sample image pairs. Each sample image pair includes a sample floating medical image and a sample reference medical image. The sample floating medical image is similar to the target floating medical image and refers to the image that needs to be transformed. The sample reference medical image is similar to the target reference medical image and refers to the image that provides a reference for the transformation of the sample floating medical image. By inputting the sample reference medical image and the target reference medical image into the medical image registration model, a predicted sample deformation field can be obtained. The sample deformation field is applied to the sample floating medical image to obtain a sample registered medical image. A first loss function can be constructed based on the sample reference medical image and the sample registered medical image. A second loss function can be constructed based on the sample deformation field. The medical image registration model is trained based on the first loss function and the second loss function, thereby improving the accuracy of the trained medical image registration model.

[0091] In a specific example, a dataset to be registered is selected and preprocessed, and then simultaneously divided into a training set and a test set. The medical images used in this example come from seven public datasets, namely OASIS, ABIDE, ADHD200, MCIC, HABS, Harvard GSP, and PPMI datasets. 3,000 images are selected as the dataset to be registered, and divided into a training set and a test set according to a certain ratio, with 80% of the data used as training data and 20% of the data used as test data, that is, 2,400 images as the training set and 600 images as the test set. During training, one image is randomly selected from the training set as a sample floating medical image, and another image is randomly selected as a sample reference medical image. The sample floating medical image and the sample reference medical image are spliced ​​in the depth direction as input to construct a sample image pair.

[0092] It should be noted that before model training and testing, the sample image pairs can be scaled to transform the image from the original size to 256x256. At the same time, during the training phase, data augmentation operations such as random horizontal flipping and random rotation can be continued on the sample image pairs.

[0093] In one embodiment, referring to Figure 6 The process of training the medical image registration model based on the first loss function and the second loss function includes:

[0094] Step 601, calculating the mean square error between the sample reference medical image and the sample registered medical image to obtain a first loss function;

[0095] Step 602: Calculate the square of the difference between each element in the sample deformation field and the corresponding element in the identical deformation field to obtain a second sub-loss function, and sum the second sub-loss functions of all elements to obtain a second loss function;

[0096] Step 603: Perform a weighted sum of the first loss function and the second loss function to obtain a target loss function;

[0097] Step 604 : Adjust parameters of the medical image registration model according to the target loss function.

[0098] In step 601 of one embodiment, the first loss function is as shown in formula (1):

[0099]

[0100] In formula (1), L mse Refers to the first loss function, f is the sample reference medical image, m is the sample floating medical image, is the output sample deformation field, is the value of the i-th element of the sample deformation field, The sample registered medical image is obtained by applying the sample deformation field to the sample reference floating medical image, p i is the coordinate of the i-th pixel, f(p i ) is the coordinate of the i-th pixel of the sample reference medical image, The coordinates of the i-th pixel of the sample registration medical image, MN refers to the upper limit of the sum.

[0101] In step 602 of one embodiment, the process of calculating the second loss function is shown in formula (2):

[0102]

[0103] In formula (2), L dfrefers to the second loss function, is the value of the i-th element of the sample deformation field, and MN is the upper limit of the summation.

[0104] It should be noted that in the sample deformation field, the element in the i-th row and j-th column represents the position of the pixel in the i-th row and j-th column of the sample registered medical image relative to that position in the sample reference floating image. For example, if it is (0,0), then the position of the pixel in the i-th row and j-th column of the sample registered medical image relative to that position in the sample floating medical image remains unchanged, that is, the pixel at this point is equal to the pixel at that point in the sample floating medical image. If it is (0.5,0.5), then the position of the pixel in the i-th row and j-th column of the sample registered medical image in the sample floating medical image is (i+0.5,j+0.5). Since it is not an integer, it can be evaluated by interpolation. At the same time, it is not difficult to find that the identity deformation field in this framework is an all-zero matrix. Therefore, formula (2) represents the distance between the sample deformation field and the identity deformation field. In fact, each element is reduced by a zero. In formula (2), it represents the elements in the vector that flattens the matrix.

[0105] In step 603 of one embodiment, the process of calculating the target loss function is shown in formula (3):

[0106] L all =a*L mse +b*L df Formula (3).

[0107] In formula (3), L all is the target loss function, a is the first weight corresponding to the first loss function, and b is the second weight corresponding to the second loss function. Here, a + b = 1. For example, a = 0.9 and b = 1.

[0108] It should be noted that L mse It measures the distance between the sample floating medical image and the sample reference medical image after the deformation field is applied. df The first and second weights measure the distance between the sample deformation field and the identical deformation field, that is, the degree of closeness between the deformation field mapping and the identity mapping. The first and second weights are effectively a trade-off. In image registration, the predicted deformation field is expected to be as close as possible to the identical deformation field, in accordance with the principle of least action. At the same time, the distance between the sample floating medical image and the sample reference medical image after the deformation field is applied is expected to be as close as possible. However, there is a conflict between these two considerations. Therefore, the first and second weights actually reflect the degree of attention paid to the two types of losses. To this end, Bayesian optimization can be used to determine the weights to obtain the first and second weights.

[0109] In step 604 of one embodiment, a loss threshold may be preset. When the target loss function is less than the loss threshold, the training process ends. When the target loss function is greater than or equal to the loss threshold, the parameters of the medical image registration model are adjusted until the target loss function is less than the loss threshold.

[0110] The advantage of the embodiment of steps 601 to 604 is that the medical image registration model is trained based on the constructed target loss function, thereby improving the accuracy of the trained medical image registration model.

[0111] In one embodiment, the candidate registration model with the highest evaluation index on the test data is determined as the medical image registration model. Multiple registration models are evaluated on the test data, and the value of the Dice evaluation index is calculated. The expression of the Dice evaluation index is as follows:

[0112]

[0113] In formula (4), the meanings of the symbols in the Dice index are as follows: |X| represents the number of pixels in the label image of the sample registered medical image obtained after sample floating medical image registration, |Y| represents the number of pixels in the label image of the sample reference medical image, and |X∩Y| represents the number of pixels in the overlapping area of ​​the two label images belonging to the same tissue at the same spatial coordinates. When the Dice index value of the model on the test data reaches the highest value, it proves that the model has the best registration effect at this time, and this candidate registration model is selected as the final medical image registration model.

[0114] In practical applications, the medical image registration model can be trained on a workbench equipped with an NVIDIA-A100 GPU. The programming language is Python 3.11, and the Pytorch 2.0 deep learning framework is used. The batch size of the training phase is set to 8, the optimizer is Adam, and a total of 150 rounds of training are performed. The initial learning rate is set to 10 -4 , using the “Poly” strategy to adjust the learning rate.

[0115] Figure 7 This is a diagram showing the registration effect of the medical image registration model of the embodiment of the present application on the human brain OCT image. Figure 7 As shown, from left to right are visualization images of a sample floating medical image, a sample reference medical image, a sample registered medical image, and a sample deformation field.

[0116] The values ​​of the indexes of different tissue registration when the Dice index of the medical image registration model on the test data is the highest are shown in Table 1:

[0117] Table 1

[0118]

[0119]

[0120] Through experiments and statistics, the medical image registration model of the present application has high performance indicators in image registration and has high registration accuracy.

[0121] In step 103 of some embodiments, the encoding module is a cascade of multiple encoding modules, and the target encoding feature includes the encoding output feature of each encoding module. In one embodiment, step 103 includes: performing image stitching on the target floating medical image and the target reference medical image to obtain a target stitched medical image; determining the encoding input feature of the first encoding module in the cascade of multiple encoding modules based on the target stitched medical image, so that the encoding modules after the first encoding module use the encoding output feature of the previous encoding module after the maximum pooling process as the encoding input feature, and each encoding module performs feature encoding based on the encoding input feature to obtain the encoding output feature.

[0122] It should be noted that the maximum pooling process is used to reduce the feature dimension of the encoding module output, which effectively reduces the network parameters and prevents overfitting.

[0123] In one embodiment, the encoding module includes a third-order transform convolution submodule, a convolutional attention submodule, and a fusion convolution submodule. Each encoding module performs feature encoding based on input features to obtain encoded output features, including: using the third-order transform convolution submodule to perform transform convolution on the input features to obtain a first encoded subfeature; using the convolutional attention submodule to perform attention convolution on the first encoded subfeature to obtain a second encoded subfeature; and using the fusion convolution submodule to perform dilated convolution on the second encoded subfeature to obtain an encoded output feature. This can improve the accuracy of feature encoding.

[0124] In step 104 of some embodiments, the target coding feature is input into a third-order transform convolution submodule for feature transformation to obtain a target transformation feature, and the third-order transform convolution submodule includes a third-order transform operator layer, a first convolution layer, a normalization layer, and an activation function layer. In one embodiment, step 104 specifically includes: using the third-order transform operator layer to perform a third-order feature transformation on the target coding feature to obtain a first transformation feature; using the first convolution layer to convolve the first transformation feature to obtain a second transformation feature; using the normalization layer to normalize the second transformation feature to obtain a third transformation feature; using the activation function layer to perform nonlinear processing on the third transformation feature to obtain a target transformation feature. Since the explanation of the third-order transform convolution submodule has been detailed in the description of step 102, it will not be repeated here. The third-order transform convolution submodule set in the above embodiment helps to improve the accuracy of medical image registration.

[0125] In step 105 of some embodiments, the target transformation feature is input into the convolutional attention module for feature extraction to obtain the target attention feature, and the convolutional attention submodule includes a second convolutional layer, an attention layer, an upsampling layer, and a residual layer. In one embodiment, step 105 specifically includes: using the second convolutional layer to convolve the target transformation feature to obtain a first extracted feature; using the attention layer to perform attention processing on the first interaction feature to obtain a second extracted feature; using the upsampling layer to upsample the second interaction feature to obtain a third extracted feature; using the residual layer to splice the target transformation feature and the third extracted feature to obtain the target attention feature. Since the explanation of the convolutional attention submodule has been detailed in the description of step 102, it will not be repeated here. The convolutional attention submodule set in the above embodiment helps to improve the accuracy of medical image registration.

[0126] In step 106 of some embodiments, the target attention feature is input into the fusion convolution submodule for feature convolution to obtain the target fusion feature, and the fusion convolution submodule includes a plurality of cascaded dilated convolution layers. In one embodiment, step 106 specifically includes: using the target attention feature as the convolution input feature of the cascaded multiple dilated convolution layers, so that each dilated convolution layer after the first dilated convolution layer uses the convolution output feature of the previous dilated convolution layer as the convolution input feature, and each dilated convolution layer performs dilated convolution on the convolution input feature to obtain the convolution output feature; and using the convolution output feature output by the last dilated convolution layer as the target fusion feature. Since the explanation of the fusion convolution submodule has been detailed in the description of step 102, it will not be repeated here. The fusion convolution module set in the above embodiment helps to improve the accuracy of medical image registration.

[0127] In step 107 of some embodiments, the decoding module is a plurality of cascaded decoding submodules, and one decoding module corresponds to one encoding module. Specifically, each cascaded decoding module corresponds to each cascaded encoding submodule in reverse order. Figure 8 The decoding module is a cascade of four decoding modules, and the encoding module is a cascade of four decoding modules. The first decoding module corresponds to the fourth encoding module. The second decoding module corresponds to the third encoding module. The third decoding module corresponds to the second encoding module. The fourth decoding module corresponds to the first encoding module.

[0128] In one embodiment, step 107 specifically includes: using the output features of the last encoding module and the target attention features as the decoding input features of the first decoding module in the cascade of multiple decoding modules, so that each decoding module after the first decoding module uses the decoding output features of the previous decoding module and the encoding output features of the encoding module corresponding to the decoding module as decoding input features, and each decoding module performs feature decoding according to the decoding input features to obtain decoding output features; and determines the target deformation field according to the decoding output features of the last decoding sub-module.

[0129] In one embodiment, each decoding module includes an upsampling submodule, a third-order transform convolution submodule, a convolutional attention submodule, and a fusion convolution submodule. Each decoding module performs feature decoding based on the decoded input features to obtain decoded output features, including: upsampling the decoded output features of the previous decoding submodule using the upsampling submodule, concatenating the upsampled features with the encoded output features of the encoding module corresponding to the decoding module to obtain a first decoded subfeature; performing transform convolution on the first decoded subfeature using the third-order transform convolution submodule to obtain a second decoded subfeature; performing attention convolution on the second decoded subfeature using the convolutional attention submodule to obtain a third decoded subfeature; and performing dilated convolution on the third decoded subfeature using the fusion convolution submodule to obtain a decoded output feature. This improves the accuracy of feature decoding, thereby improving the accuracy of image registration.

[0130] In one example, referring to Figure 9 , the medical image registration model includes four cascaded encoding modules, a third-order transform convolution submodule, a convolution attention submodule, a fusion convolution submodule, and four cascaded decoding modules. Each encoding module includes a third-order transform convolution submodule, a convolution attention submodule, and a fusion convolution submodule, and each encoding submodule is followed by a pooling submodule. Each decoding module includes an upsampling submodule, a third-order transform convolution submodule, a convolution attention submodule, and a fusion convolution submodule. After each convolution attention submodule, there will be a Dropout layer to suppress overfitting. The image obtained by splicing the sample floating medical image and the sample reference medical image ( Figure 9 The image shown in Figure 2 is input into the medical image registration model to obtain the predicted sample deformation field ( Figure 9 Output shown).

[0131] Specifically, for the first encoding module, the image with 2 channels is input into the third-order transform convolution submodule to obtain the first encoding sub-feature with 16 channels. This is then passed through the convolution attention submodule to obtain the second encoding sub-feature with 16 channels. This is then passed through the dropout layer (Droupout0.3) and the fusion convolution submodule to obtain the encoding output feature with 16 channels. The pooling submodule performs maximum pooling on the encoding output feature with 16 channels to obtain the encoding input feature with 16 channels. For the second encoding module, the encoding input feature with 16 channels is input into the third-order transform convolution submodule to obtain the first encoding sub-feature with 32 channels. This is then passed through the convolution attention submodule to obtain the second encoding sub-feature with 32 channels. This is then passed through the dropout layer (Droupout0.3) and the fusion convolution submodule to obtain the encoding output feature with 32 channels. Similarly, for the third encoding module, an encoding output feature with 64 channels can be obtained. For the fourth encoding module, a coded output feature with 128 channels is obtained. The Dropout layer in the third and fourth encoding modules uses Dropout 0.2. The pooling submodule performs max pooling on the coded output feature with 128 channels to obtain the target coded feature with 128 channels.

[0132] The target encoding feature with a channel number of 128 is input into the third-order transformation convolution submodule to obtain the target transformation feature with a channel number of 256; then it passes through the convolution attention submodule to obtain the target attention feature with a channel number of 256; then it passes through the Dropout layer (Dropout0.1) and then through the fusion convolution submodule to obtain the target fusion feature with a channel number of 256.

[0133] For the first decoding module, the upsampling submodule is used to upsample the target fusion feature with a channel number of 256 to obtain the upsampled feature with a channel number of 256, and the upsampled feature is spliced ​​with the encoded output feature with a channel number of 128 to obtain the first decoding sub-feature with a channel number of 128+256; then it passes through the third-order transform convolution submodule to obtain the second decoding sub-feature with a channel number of 128; then it passes through the convolution attention submodule to obtain the third decoding sub-feature with a channel number of 128; after passing through the Dropout layer (Dropout0.2), and finally through the fusion convolution submodule, the decoding output feature with a channel number of 128 is obtained. For the second decoding module, the upsampling submodule upsamples the decoded output features with 128 channels to obtain upsampled features with 128 channels. These upsampled features are then concatenated with the encoded output features with 64 channels to obtain the first decoded sub-feature with 64 + 128 channels. This is then passed through a third-order transform convolution submodule to obtain the second decoded sub-feature with 64 channels. This is then passed through a convolutional attention submodule to obtain the third decoded sub-feature with 64 channels. This is then passed through a dropout layer (Dropout 0.2) and finally through a fusion convolution submodule to obtain the decoded output features with 64 channels. Similarly, for the third decoding module, a decoded output feature with 32 channels is obtained, and for the fourth decoding module, a decoded output feature with 16 channels is obtained. The dropout layer in the third and fourth decoding modules uses Dropout 0.3. This is then passed through a 1*1 convolutional layer to obtain a feature with 1 channel, i.e., the predicted sample deformation field. Finally, the sample deformation field is applied to the sample floating medical image to obtain the sample registered medical image.

[0134] It should be noted that Figure 9 Dropout0.1, Dropout0.2, and Dropout0.3 in the figure respectively mean that a neuron in the medical image registration model stops working with probabilities of 0.1, 0.2, and 0.3.

[0135] See also Figure 10 The present application also provides a medical image registration device that can implement the above-mentioned medical image registration method. Figure 10 This is a block diagram of the module structure of the medical image registration device provided in an embodiment of the present application, which includes:

[0136] An image acquisition unit 1001 is configured to acquire a target floating medical image and a target reference medical image;

[0137] The model acquisition unit 1002 is used to acquire a medical image registration model, which includes an encoding module, a third-order transform convolution module, a convolution attention module, a fusion convolution module, and a decoding module;

[0138] A feature encoding unit 1003 is configured to input the target floating medical image and the target reference medical image into an encoding module for feature encoding to obtain target encoding features;

[0139] The feature conversion unit 1004 is used to input the target coding feature into the third-order transform convolution module for feature conversion to obtain the target transformed feature;

[0140] The feature extraction unit 1005 is used to input the target transformation feature into the convolution attention module for feature extraction to obtain the target attention feature;

[0141] The feature convolution unit 1006 is used to input the target attention feature into the fusion convolution module for feature convolution to obtain the target fusion feature;

[0142] The feature decoding unit 1007 is used to input the target coding feature and the target fusion feature into the decoding module for feature decoding to obtain the target deformation field, and apply the target deformation field to the target floating medical image to obtain the target registered medical image.

[0143] It should be noted that the specific implementation of the medical image registration device is substantially the same as the specific embodiment of the above-mentioned medical image registration method, and will not be described in detail here.

[0144] The present application also provides an electronic device comprising: a memory, a processor, a program stored in the memory and executable on the processor, and a data bus for enabling communication between the processor and the memory. When the program is executed by the processor, the above-described medical image registration method is implemented. The electronic device can be any intelligent terminal, including a tablet computer and an in-vehicle computer.

[0145] See also Figure 11 , Figure 11 The hardware structure of an electronic device according to another embodiment is shown. The electronic device includes:

[0146] The processor 1101 can be implemented as a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application.

[0147] The memory 1102 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 1102 can store an operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1102 and is called by the processor 1101 to execute the medical image registration method of the embodiments of this application.

[0148] Input / output interface 1103, used to implement information input and output;

[0149] Communication interface 1104, used to implement communication interaction between this device and other devices, which can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WiFi, Bluetooth, etc.);

[0150] Bus 1105 , which transmits information between various components of the device (e.g., processor 1101 , memory 1102 , input / output interface 1103 , and communication interface 1104 );

[0151] The processor 1101 , the memory 1102 , the input / output interface 1103 and the communication interface 1104 are connected to each other in communication within the device via a bus 1105 .

[0152] An embodiment of the present application also provides a storage medium, which is a computer-readable storage medium used for computer-readable storage. The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the above-mentioned medical image registration method.

[0153] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0154] The embodiments described in the embodiments of this application are intended to more clearly illustrate the technical solutions of the embodiments of this application and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those skilled in the art will appreciate that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems.

[0155] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present application, and may include more or fewer steps than shown in the figures, or a combination of certain steps, or different steps.

[0156] Those skilled in the art will appreciate that all or some of the steps in the methods, systems, and functional modules / units in the devices disclosed above may be implemented as software, firmware, hardware, or appropriate combinations thereof.

[0157] The terms "first", "second", "third", "fourth", etc. (if any) in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential order. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0158] It should be understood that in this application, "at least one (item)" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following items" or similar expressions refers to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b or c can mean: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, c can be single or multiple.

[0159] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0160] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0161] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes multiple instructions for enabling an electronic device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store programs, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0162] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and essence of the present invention should be within the scope of the present invention.

Claims

1. A medical image registration method, characterized in that: The method comprises: acquiring a target floating medical image and a target reference medical image; Acquire a medical image registration model, wherein the medical image registration model includes an encoding module, a third-order transform convolution submodule, a convolution attention submodule, a fusion convolution submodule, and a decoding module; Inputting the target floating medical image and the target reference medical image into the encoding module for feature encoding to obtain target encoding features; Inputting the target coding feature into the third-order transform convolution submodule for feature transformation to obtain the target transformed feature; Inputting the target transformation feature into the convolutional attention submodule for feature extraction to obtain the target attention feature; Inputting the target attention feature into the fusion convolution submodule for feature convolution to obtain the target fusion feature; Inputting the target coding feature and the target fusion feature into the decoding module for feature decoding to obtain a target deformation field, and applying the target deformation field to the target floating medical image to obtain a target registered medical image; In which, the medical image registration model is trained based on the first loss function and the second loss function of each sample image pair in the sample image set, the sample image pair includes a sample floating medical image and a sample reference medical image, the first loss function is constructed based on the distance between the sample reference medical image and the sample registered medical image, and the second loss function is constructed based on the distance between the sample deformation field and the preset identical deformation field. The sample deformation field is obtained by inputting the sample floating medical image and the sample reference medical image into the medical image registration model, and the sample registered medical image is obtained by applying the sample deformation field to the sample floating medical image, and each element of the identical deformation field is zero.

2. The method according to claim 1, characterized in that The third-order transform convolution submodule includes a third-order transform operator layer, a first convolution layer, a normalization layer, and an activation function layer; The step of inputting the target coding feature into the third-order transform convolution submodule for feature transformation to obtain the target transformed feature includes: Using the third-order transformation operator layer, performing a third-order feature transformation on the target coding feature to obtain a first transformation feature; Convolving the first transformed features using the first convolutional layer to obtain second transformed features; Normalizing the second transformed features using the normalization layer to obtain third transformed features; The activation function layer is used to perform nonlinear processing on the third transformation feature to obtain the target transformation feature.

3. The method according to claim 1, characterized in that The convolutional attention submodule includes a second convolutional layer, an attention layer, an upsampling layer, and a residual layer; The step of inputting the target transformation feature into the convolutional attention submodule for feature extraction to obtain the target attention feature comprises: Using the second convolutional layer, convolving the target transformation feature to obtain a first extracted feature; Using the attention layer, performing attention processing on the first extracted features to obtain second extracted features; Upsampling the second extracted features using the upsampling layer to obtain third extracted features; The target transformation feature and the third extracted feature are concatenated using the residual layer to obtain the target attention feature.

4. The method according to claim 1, wherein The fused convolution submodule includes a plurality of cascaded dilated convolution layers; Inputting the target attention feature into the fusion convolution submodule to obtain the target fusion feature includes: The target attention feature is used as the convolution input feature of multiple cascaded dilated convolutional layers, so that each dilated convolutional layer after the first dilated convolutional layer uses the convolution output feature of the previous dilated convolutional layer as the convolution input feature, and each dilated convolutional layer performs dilated convolution on the convolution input feature to obtain the convolution output feature; The convolution output feature output by the last spatiotemporal convolution layer is used as the target fusion feature.

5. The method according to claim 1, wherein The encoding module is a plurality of cascaded encoding modules, and the target encoding feature includes the encoding output feature of each encoding module; The step of inputting the target floating medical image and the target reference medical image into the encoding module for feature encoding to obtain target encoding features includes: performing image stitching on the target floating medical image and the target reference medical image to obtain a target stitched medical image; The encoding input features of the first encoding module in the cascaded multiple encoding modules are determined according to the target spliced ​​medical image, so that each encoding module after the first encoding module uses the encoding output features of the previous encoding module after maximum pooling processing as the encoding input features, and each encoding module performs feature encoding according to the encoding input features to obtain the encoding output features.

6. The method according to claim 5, characterized in that The decoding module is a plurality of cascaded decoding modules, and one decoding module corresponds to one encoding module; Inputting the target encoding feature and the target attention feature into the decoding module for feature decoding to obtain a target deformation field, including: Perform feature splicing on the output feature of the last encoding module and the target attention feature to obtain a target splicing feature; Using the target splicing features as decoding input features of a plurality of cascaded decoding modules, so that each decoding module after the first decoding module uses the decoding output features of the previous decoding module after upsampling processing and the encoding output features of the encoding module corresponding to the decoding module as decoding input features, and each decoding module performs feature decoding according to the decoding input features to obtain the decoding output features; The target deformation field is determined according to the decoding output feature of the last decoding module.

7. The method according to any one of claims 1 to 6, characterized in that The medical image registration model is trained based on a first loss function and a second loss function for each sample image pair in the sample image set, including: Calculating a mean square error between the sample reference medical image and the sample registered medical image to obtain a first loss function; Calculate the square of the difference between each element in the sample deformation field and the corresponding element in the identity deformation field to obtain a second sub-loss function, and sum the second sub-loss functions of all elements to obtain a second loss function; Obtain a target loss function by performing a weighted summation on the first loss function and the second loss function; Parameters of the medical image registration model are adjusted according to the target loss function.

8. A medical image registration device, characterized in that: The device comprises: an image acquisition unit, configured to acquire a target floating medical image and a target reference medical image; A model acquisition unit, configured to acquire a medical image registration model, wherein the medical image registration model comprises an encoding module, a third-order transform convolution submodule, a convolution attention submodule, a fusion convolution submodule, and a decoding module; A feature encoding unit, configured to input the target floating medical image and the target reference medical image into the encoding module for feature encoding to obtain target encoding features; A feature conversion unit, configured to input the target coding feature into the third-order transform convolution submodule for feature conversion to obtain a target transformed feature; A feature extraction unit, configured to input the target transformation feature into the convolutional attention submodule for feature extraction to obtain a target attention feature; A feature convolution unit, configured to input the target attention feature into the fusion convolution submodule for feature convolution to obtain a target fusion feature; a feature decoding unit, configured to input the target coding feature and the target fusion feature into the decoding module for feature decoding to obtain a target deformation field, and apply the target deformation field to the target floating medical image to obtain a target registered medical image; In which, the medical image registration model is trained based on the first loss function and the second loss function of each sample image pair in the sample image set, the sample image pair includes a sample floating medical image and a sample reference medical image, the first loss function is constructed based on the distance between the sample reference medical image and the sample registered medical image, and the second loss function is constructed based on the distance between the sample deformation field and the preset identical deformation field. The sample deformation field is obtained by inputting the sample floating medical image and the sample reference medical image into the medical image registration model, and the sample registered medical image is obtained by applying the sample deformation field to the sample floating medical image, and each element of the identical deformation field is zero.

9. An electronic device, characterized in that: The electronic device includes a memory and a processor, the memory stores a computer program, and the processor implements the method according to any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images